LendMatch: An AI-Powered Holistic Credit Risk Platform with Dual-Perspective Architecture, Secure Ledger, Multilingual Support, and Integrated Chatbot for Intelligent Loan Underwriting

Traditional credit risk analysis in banking often tends to neglect the 'new-to-credit' population due to their lack of historical credit scoring. In the conventional approach, people who do not have any credit history but can still afford to pay back their loans are often neglected, such as gig-economy employees, rural inhabitants, and young professionals. To tackle the problem and make financial services available to the 'new-to-credit' population, this paper proposes the use of a holistic AI-driven credit risk evaluation tool called LendMatch. LendMatch utilizes alternative underwriting data rather than historical data for evaluating potential loan applicants.

Instead of analyzing historical behavior, this project suggests evaluating an applicant's eligibility for receiving a loan via proxy labels created using deterministic calculations of two proxies: Earning Power and Asset Coverage. Using risk tier thresholds based on Loan-to-Income (LTI) ratio and Asset-to-Loan ratio, the system creates synthetic proxy labels that simulate real-world expert-level decision-making policies. The project uses an eXtreme Gradient Boosting (XGBoost) classifier supplemented by the Synthetic Minority Over-sampling Technique (SMOTE).

Empirical experiments show that the suggested XGBoost algorithm successfully applies the strict criteria embedded into underwriting algorithms to predict applicants' eligibility for receiving a loan. Besides the core classifier component, LendMatch includes several additional features that target both sides of the lending process, including Explainable AI (XAI) in multiple languages to justify decisions on loan applications based on an asset portfolio, an institution's matchmaker, a database in SQLite format for keeping track of decisions, and a chatbot.